Download A gentle introduction to von Neumann algebras for model theorists

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

History of algebra wikipedia , lookup

Complexification (Lie group) wikipedia , lookup

Clifford algebra wikipedia , lookup

Fundamental theorem of algebra wikipedia , lookup

Congruence lattice problem wikipedia , lookup

Oscillator representation wikipedia , lookup

Transcript
A GENTLE INTRODUCTION TO VON NEUMANN
ALGEBRAS FOR MODEL THEORISTS
ISAAC GOLDBRING
Contents
1. Topologies on B(H) and the double commutant theorem
2. Examples of von Neumann algebras
2.1. Abelian von Neumann algebras
2.2. Group von Neumann algebras and representation theory
2.3. The Hyperfinite II1 factor R
3. Projections, Type Classification, and Traces
3.1. Projections and the spectral theorem
3.2. Type classification of factors
3.3. Traces
4. Tracial ultrapowers and the Connes Embedding Problem
4.1. The 2-norm
4.2. Tracial ultraproducts
4.3. The Connes Embedding Problem
5. Some Model Theory of tracial von Neumann algebras
5.1. Axiomatizability
5.2. The model theoretic version of CEP
5.3. Model companions and connection to CEP
5.4. Stability
References
2
5
6
6
9
10
10
12
13
15
15
16
17
17
18
20
20
22
23
These notes were part of my course on Continuous Model Theory at UIC in
the fall of 2012. We spent several weeks on the model theory of von Neumann
algebras and these notes served as an introduction to von Neumann algebras
with the intended audience being model theorists who know a little bit of
functional analysis. These notes borrow in great part from excellent notes
of Vaughn Jones [10], Martino Lupini and Asger Tornquist [11], and Jacob
Lurie [12].
Date: May 3, 2013.
1
2
ISAAC GOLDBRING
1. Topologies on B(H) and the double commutant theorem
In this section, H denotes a (complex) Hilbert space and B(H) denotes the
set of bounded operators on H: recall that the linear operator T : H → H
is bounded if T (B1 (H)) is bounded, where B1 (H) := {x ∈ H : kxk ≤ 1}
is the closed unit ball in H. It is straightforward to check that the bounded
operators are precisely the continuous linear operators.
For T ∈ B(H), we set kT k := sup{kT xk : kxk ≤ 1}, and call kT k the
(operator) norm of T . It is easy to see that kT k is the radius of the smallest
closed ball centered at 0 in H containing T (B1 (H)). As the name indicates,
k · k is a norm on B(H) and we call the resulting topology the norm topology
on B(H).
Exercise 1.1. For T ∈ B(H) and v ∈ H, we have kT vk ≤ kT k · kvk.
Recall that for each T ∈ B(H), there is a unique function T ∗ : H → H
satisfyiing hT x, yi = hx, T ∗ yi for all x, y ∈ H. T ∗ is called the adjoint of T
and plays a crucial role in operator algebras. It is easy to see that T ∗ is once
again in B(H) and that (T ∗ )∗ = T .
We now introduce two more topologies on B(H). First, the strong (operator) topology on B(H) is defined to be the topology on B(H) whose subbasic
open neighborhoods are of the form
UT0 ,v, := {T ∈ B(H) : kT v − T0 vk < },
as T0 , v and range over B(H), H, and R>0 respectively.
Exercise 1.2. The strong topology is the weakest topology on B(H) that
makes, for each v ∈ H, the map T 7→ kT vk : B(H) → R continuous.
The weak (operator) topology on B(H) is the topology on B(H) whose
subbasic open neighborhoods are of the form
VT0 ,v,w, := {T ∈ B(H)| : |hT v, wi − hT0 v, wi| < ,
as T0 , (v, w), and > 0 range over B(H), H × H, and R>0 respectively.
Exercise 1.3. The weak topology is the weakest topology on B(H) that
makes, for each v, w ∈ H, the map T 7→ hT v, wi : B(H) → R continuous.
Let ONT, SOT, and WOT denote the operator norm, strong, and weak
topologies respectively.
Lemma 1.4. WOT ⊆ SOT ⊆ ONT.
Proof. First suppose that OT∈ WOT. Fix T0 ∈ O. Fix v1 , . . . , vn , w1 , . . . , wn ∈
H and > 0 such that ni=1 VT0 ,vi ,wi , ⊆ O. Without loss of generality, we may assume that wi 6= 0 for each i. By Cauchy-Schwarz, we have
|h(T − T0 )vi , wi i| ≤ k(T − T0 )vi k · kwi k; it follows that
n
n
\
\
UT0 ,vi , kw k ⊆
VT0 ,vi ,wi , ⊆ O.
i=1
i
i=1
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
3
Thus, O ∈ SOT.
Similarly, suppose
T that O ∈ WOT and T0 ∈ O. Fix v1 , . . . , vn ∈ H and
> 0 such that ni=1 UT0 ,vi , ⊆ O. Without loss of generality, each vi 6= 0.
Since k(T − T0 )vi k ≤ kT − T0 kkvk, it follows that
Br (T0 ) ⊆
n
\
UT0 ,vi , ⊆ O,
i=1
where r := min1≤i≤n kvi k and Br (T0 ) := {T ∈ B(H) : kT − T0 k < r}. It
follows that O ∈ ONT.
Why consider other topologies on B(H) other than ONT? Consider the
following example:
Example 1.5. Let `2 be the separable Hilbert space, that is,
X
`2 = {x = (xn ) ∈ CN : kxk2 :=
|xn |2 < ∞}
n
P
with inner product hx, yi := n xn yn . Let L ∈ B(H) be the left-shift operator, that is, L(x) = (0, x0 , x1 , . . .). It is easy to see that L∗ = R, the rightshift operator given by R(x) := (x1 , x2 , . . .). Notice that P
Li → 0 as i → ∞
2
i
2
in the strong topology as, for x ∈ ` , we have kL xk = ∞
n=i |xi | → 0 as
i → ∞ since v ∈ `2 . On the other hand, (Li )∗ = Ri 6→ 0 as i → ∞ as, for
x = (1, 0, 0, . . .), we have kRi xk = 1 for all i. Thus, we see that the map
T 7→ T ∗ : B(`2 ) → B(`2 ) is not strongly continuous, that is, not continuous
with respect to SOT.
Contrast the previous example with the following lemma:
Lemma 1.6. The map T 7→ T ∗ : B(H) → B(H) is weakly continuous, that
is, is continuous with respect to WOT.
Proof. The inverse image of UT0 ,v,w, under the adjoint map is UT0∗ ,w,v, . Exercise 1.7. For any S0 , T0 ∈ B(H), the maps
T 7→ S0 T, S 7→ ST0 : B(H) → B(H)
are both weakly continuous; that is, composition in B(H) is separately weakly
continuous. (Extra credit: Show that composition itself need not be jointly
weakly continuous in general.)
wk
st
For B ⊆ B(H), let B (resp. B ) denote the closure of B with respect
to WOT (resp. SOT).
We call B ⊆ B(H) a subalgebra of B(H) if B is closed under scalar multiplication, addition, and composition. If B is also closed under taking adjoint,
we call B a ∗-subalgebra of B(H). If the identity operator I belongs to the
subalgebra B, we say that B is a unital subalgebra of B(H).
Lemma 1.6 and Exercise 1.7 are enough to prove the following important
result:
4
ISAAC GOLDBRING
Proposition 1.8. If A ⊆ B(H) is a ∗-subalgebra of B(H), then so is A
X0
wk
.
X0
For X ⊆ B(H), set
:= {T ∈ B(H) : ST = T S for all S ∈ X}.
is
called the commutant of X. We set X 00 := (X 0 )0 , the double commutant of
X.
Exercise 1.9. Suppose that X ⊆ B(H).
(1) X 0 is a weakly closed subalgebra of B(H).
(2) If X is closed under adjoints, then X 0 is a weakly closed ∗-subalgebra
of B(H).
wk
(3) X ⊆ X 00 .
(4) X 0 = X 000 := (X 0 )00 .
The following theorem can be considered the beginning of von Neumann
algebra theory.
Theorem 1.10 (von Neumann double commutant theorem). Suppose that
A ⊆ B(H) is a unital ∗-subalgebra of B(H). Then A is strongly (and hence
weakly) dense in A00 .
Before we prove Theorem 1.10, let us state the following important corollaries.
st
Corollary 1.11. For A as in Theorem 1.10, we have A = A
wk
= A00 .
Proof. By Exercise 1.9(3) (and the definition of the topologies), we have
st
A ⊆A
st
By Theorem 1.10, we have A00 ⊆ A .
wk
⊆ A00 .
Corollary 1.12. For A ⊆ B(H) a unital ∗-subalgebra of B(H), the following
are equivalent:
(1) A = X 0 for some X ⊆ B(H);
(2) A = A00 ;
(3) A is weakly closed;
(4) A is strongly closed.
Proof. (1) ⇒ (2) follows from Exercise 1.9(4). (2) ⇒ (3) ⇒ (4) ⇒ (1) follows
from the previous corollary.
Definition 1.13. A unital ∗-subalgebra of B(H) satisfying any of the equivalent conditions of Corollary 1.12 is called a von Neumann algebra.
What makes von Neumann algebras such a robust notion is the equivalence of the algebraic conditions in (1) and (2) of Corollary 1.12 with the
topological conditions (3) and (4).
Proof of Theorem 1.10. Fix T0 ∈ A00 ; we must show that any basic strongly
open neighborhood of T0 intersects A. We first deal with the special case
of a subbasic strongly open neighborhood of T ; the general case of a basic
strongly open neighborhood will follow from an “amplification” trick.
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
5
We thus fix v ∈ H and > 0; we aim to prove that UT0 ,v, ∩ A 6= ∅. Set
A · v := {T v : T ∈ A} and let P : H → H denote the orthogonal projection
onto the closed subspace A · v of H. Notice that v ∈ A · v (since A is unital)
and that A · v is A-invariant, that is, T (A · v) ⊆ A · v for each T ∈ A. We
claim that P ∈ A0 . Towards this end, first observe that if w ∈ (A · v)⊥ and
S, T ∈ A, we have hSw, T vi = hw, S ∗ T V i = 0 since S ∗ T v ∈ A · v (here we
used that A is a ∗-subalgebra). It follows that (A · v)⊥ is A-invariant. Thus,
for T ∈ A and w ∈ H, we have
P T w = P T P w + P T (I − P )w = P T P w = T P w;
since w ∈ H was arbitrary, this shows that P T = T P , and consequently,
P ∈ A0 , as desired. Now, since T0 ∈ A00 , we have that T0 and P commute,
whence T0 v = T0 P v = P T0 v, that is, T0 v ∈ A · v. Thus, there is T ∈ A such
that kT0 v − T vk < , that is, T ∈ UT0 ,v, ∩ A.
For the general case, we now fix v1 , . . . , vn ∈ H and > 0; we aimL
to prove
that UT0 ,v1 ,...,vn , ∩ A 6= ∅. The idea is to pass from H to H (n) := ni=1 Hi ,
the direct sum of n many copies of H. Now, the tuple of vectors (v1 , . . . , vn )
from H is a single vector in H (n) .
Let us make a few preliminary comments about H (n) . First, given Tij ∈
B(H), i, j = 1, . . . , n, we get an operator [Tij ] ∈ B(H (n) ) given by “matrix
multiplication,”
that is, for k ∈ {1, . . . , n}, the k th component of [Tij ](h1 , . . . , hn )
Pn
is
j=1 Tkj hj . We leave it to the reader to check that every element of
(n)
B(H ) is of the form [Tij ] for a suitable choice of elements Tij from B(H).
Now, given T ∈ B(H), we consider T (n) ∈ B(H) which is the “diagonal matrix” all of whose diagonal entries are T ; formall, T (n) = [Tij ], where Tii = T
and Tij = O for i 6= j.
We now return to the proof. Set A(n) := {T (n) : T ∈ A} ⊆ B(H (n) ),
a unital ∗-subalgebra of B(H (n) ). Observe now that [Tij ] ∈ (A(n) )0 if and
(n)
only if each Tij ∈ A0 , whence T0 ∈ (A(n) )0 . By the first part of the proof,
there is T ∈ A such that T (n) ∈ UT (n) ,(v ,...,v ), ∩ A(n) . It follows that
n
1
0
T ∈ UT0 ,v1 ,...,vn , ∩ A.
Remark 1.14. A unital ∗-subalgebra of B(H) that is closed in the norm
topology is called a C ∗ -algebra. We thus see that every von Neumann algebra
is a C ∗ -algebra; the converse does not hold. We should point out that the
general theory of C ∗ -algebras and von Neumann algebras are wildly different.
If X is any subset of B(H), we call the smallest von Neumann algebra
containing X the von Neumann algebra generated by X. By Proposition 1.8,
we see that the von Neumann algebra generated by X is (X ∪ X ∗ )00 , where
X ∗ := {T ∗ : T ∈ X}.
2. Examples of von Neumann algebras
We now present a list of examples of von Neumann algebras. Of course,
for any Hilbert space H, B(H) is a von Neumann algebra; in particular, for
6
ISAAC GOLDBRING
any n ∈ N, Mn (C) is a von Neumann algebra (this is isomorphic to B(H)
when dim(H) = n). We now turn to less trivial examples.
2.1. Abelian von Neumann algebras. A von Neumann algebra A is said
to be abelian if T S = ST for all S, T ∈ A. For example, suppose that (X, µ)
is a σ-finite measure space. We set
L∞ (X, µ) := {f : X → C : f is measurable and essup f < ∞}
and
2
Z
L (X, µ) := {f : X → C : f is measurable and
|f |2 dµ < ∞}.
X
We then have an algebra embedding L∞ (X, µ) → B(L2 (X, µ) given by
f 7→ mf , where mf (g) := f g. We often identify L∞ (X, µ) with its image
under this embedding.
Exercise 2.1. L∞ (X, µ)0 = L∞ (X, µ). (Hint: In showing that L∞ (X, µ)0 ⊆
L∞ (X, µ), first assume that µ(X) < ∞. In that case, 1 ∈ L2 (X, µ); for
T ∈ L∞ (X, µ)0 , set f := T (1). Show that f ∈ L∞ (X, µ) and that T (g) = f g
for all g ∈ L∞ (X, µ), which is sufficient since L∞ (X, µ) is dense in L2 (X, µ).
The general σ-finite case reduces to the case of finite measure by considering
restrictions to suitable finite measure sets.)
Consequently, we see that L∞ (X, µ) is a von Neumann algebra. Moreover,
mf mg = mf g , we see that it is an abelian von Neumann algebra. It is a fact
that all abelian von Neumann algebras are of the form L∞ (X, µ) for some
(X, µ). It is for this reason that von Neumann algebra theory is sometimes
regarded as noncommutative measure theory.
1 equipped with its haar measure and
Example 2.2. Suppose that X = SP
∞
1
suppose that f ∈ L (S ). Let sn := nk=−n cn eikθ be the nth partial sum of
the Fourier series for f . Since sn converges to f in L2 (S1 ), we have that msn
converges to mf strongly in B(L2 (S1 )). It follows that L∞ (S1 ) is generated
(as a von Neumann algebra) by the functions eikθ for k ∈ Z.
2.2. Group von Neumann algebras and representation theory. Recall that U (H) consists of the elements U of B(H) for which U ∗ = U −1 . It
is clear that U (H) is a group (under composition) and is referred to as the
unitary group of H.
For the rest of this subsection, we suppose that Γ is an arbitrary countable
(discrete) group. A unitary group representation of Γ is a group homomorphism α : Γ → U (H). The von Neumann algebra generated by α(Γ), which
is just α(Γ)00 as α(Γ)∗ = α(Γ), is referred to as the group von Neumann
algebra of the representation α.
There is a compelling reason for studying α(Γ)00 . A major concern in
representation theory is to understand the invariant subspaces of a representation, that is, the closed subspaces H0 of H such that α(γ)(H0 ) ⊆ H0
for each γ ∈ Γ. First notice that if H0 is an invariant subspace of α and
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
7
P : H → H is the orthogonal projection onto H0 , then P ∈ α(Γ)0 . Indeed,
for γ ∈ Γ, and x = x1 + x2 ∈ H, with x1 ∈ H0 and x2 ∈ H0⊥ , we have
P α(γ)(x) = P α(γ)(x1 ) + P α(γ)(x2 ) = α(γ)(x1 ) = α(γ)(P x1 ).
In the above calculation, we used that both H0 and H0⊥ are invariant (the
latter is invariant since each α(γ) is unitary). Conversely, it is easy to see
that if P : H → H is an abstract projection, that is, P 2 = P ∗ = P , which
lies in α(Γ)0 , then P (H) is an invariant subspace of H.
It follows that an understanding of the invariant subspaces of α amounts
to an understanding of the projections of the von Neumann algebra α(Γ)0 .
Since the projections of a von Neumann algebra generate the von Neumann
algebra (we will discuss in Subsection 3.1), it follows that understanding
the invariant subspaces of α amounts to understanding α(Γ)0 , or somewhat
equivalently, understanding α(Γ)00 .
There is a particular unitary representation of Γ that is of extreme importance, the so-called left
To define this, we first set
P regular representation.
2 < ∞}. We make `2 (Γ) into a Hilbert
`2 (Γ) := {f : Γ → C :
|f
(γ)|
γ∈Γ
P
space by equipping it with the inner product hf, gi := γ∈Γ f (γ)g(γ). (This
is precisely the same thing as `2 , the only difference being that we index our
sequences by Γ rather than N.) For γ ∈ Γ, we define uγ : `2 (Γ) → `2 (Γ)
by uγ (f )(η) := f (γ −1 η). It is easy to see that uγ is a linear operator and
2
that u∗γ = u−1
γ = uγ −1 , so that each uγ ∈ U (` (Γ)). Moreover, uγ uρ = uγρ ,
whence we see that u : Γ → U (`2 (Γ)) given by u(γ) := uγ is a unitary
representation of Γ, referred to as the left regular representation. (As one
might imagine, if we had used the right action rather than the left action,
we would then be defining the right regular representation.) It is easy to
check that {uγ : γ ∈ Γ} is a linearly independent set in B(`2 (Γ)), whence
the algebra generated by the u(Γ) is isomorphic to the group ring CΓ. The
von Neumann algebra associated to the left regular representation is simply
called the group von Neumann algebra corresponding to Γ and is denoted by
L(Γ).
For γ ∈ Γ, let γ ∈ `2 (Γ) be the standard basis vector corresponding to
γ, that is, γ (η) = δγη . Then {γ : γ ∈ Γ}
P form an orthonormal basis
for `2 (Γ) and, for f ∈ `2 (Γ), we write f = γ∈Γ f (γ)γ . What does the
“matrix representation” for uγ look like with respect to the aforementioned
basis for `2 (Γ)? Well, notice that uγ (ρ ) = γρ , so the (η, ρ) entry of the
matrix for uγ is 1 if γρ = η and 0 otherwise, that is, the (η, ρ) entry of uγ is
1 if ηρ−1 = γ and 0 otherwise. Let us call an infinite matrix quasidiagonal
if the (η, ρ) entry only depends on ηρ−1 . It follows that the matrix for uγ
is quasidiagonal. Notice also that any linear combination of quasidiagonal
matrices is also quasidiagonal. Less obvious is the fact that if T ∈ L(Γ), then
the matrix for T is quasidiagonal. Indeed, the (η, ρ) entry of T is hT ρ, ηi,
which is a limit of a net of the form hTα ρ, ηi, where each Tα is in the algebra
generated by u(Γ).
8
ISAAC GOLDBRING
Based on the last observation
of the previous paragraph, we write elements
P
of L(Γ) as formal sums γ∈Γ cγ uγ , where cγ represents the constant value
of the “diagonal” ηρ−1 = γ. (Of course, if cγ 6= 0 for only finitely many γ,
that then signifies that we are looking at an element of the group algebra.)
In general, it is quite difficult to establish what sequences (cγ )γ∈Γ are the
coefficients of an element of L(Γ), but it is useful to at least observe the
following:
P
Lemma 2.3. If γ∈Γ cγ uγ represents an element of L(Γ), than the function
γ 7→ cγ belongs to `2 (Γ).
P
Proof. Notice that ( γ∈Γ cγ uγ )(id )(η) = cη ; here id denotes the identity of
P
the group. Since ( γ∈Γ cγ uγ )(id ) ∈ `2 (Γ), it follows that η 7→ cη belongs to
`2 (Γ).
Here is a case we can fully analyze:
Example
2.4. Suppose
that Γ = Z. The map V : `2 (Γ) → L2 (S1 ) defined by
P
P
ikθ
V ( k ck k ) := k ck e is an (isometric) isomorphism. Observe also that
(V ul V −1 )(eikθ ) = eilθ eikθ , that is, V ul V −1 = meilθ (since the eikθ form an
orthonormal basis for L2 (S1 )). By Example 2.2, we know that the operators
meikθ generate the von P
Neumann algebra L∞ (S1 ), and, for f ∈ L∞ (S 1 ), we
−1
have that V mf V = k ck uk , where ck = thePFourier coefficients for f .
It follows that L(Γ) consists of all formal sums k ck uk , where (ck ) is the
Fourier coefficients for an element of L∞ (S1 ).
For an arbitrary von Neumann algebra A, set
Z(A) := {T ∈ A : T S = ST for all S ∈ A},
the center of A.
Example 2.5. Suppose that Γ = Fn , the free group on n generators,
P for n ≥
2. Consider the question: what is the center of L(Γ)? Note that γ cγ uγ ∈
P
P
Z(L(Γ)) if and only if ( γ cγ uγ )uρ = uρ ( γ cγ uγ ); it is routine to check
that thisPlatter condition is equivalent to cγργ −1 = cρ for all γ, ρ ∈ Γ. In other
words, γ cγ uγ ∈ Z(L(Γ)) if and only if the function γ 7→ cγ is constant
on conjugacy classes. In our case, all nontrivial conjugacy classes, that is,
all conjugacy classes other than {id}, are infinite. Since γ 7→ cγ belongs to
`2 (Γ) (by Lemma 2.3), it follows that cγ = 0 for γ 6= id. Consequently, we
have that Z(L(Γ)) = {cuid : c ∈ C} = C · I.
The phenomenon exhibited in the previous example is important enough
to merit a definition.
Definition 2.6. A von Neumann algebra A is called a factor if Z(A) = C·I.
Observe that C · I is always contained in Z(A), so being a factor means
that the scalar operators are the only elements of the center. Observe also
that the only property of Fn used to conclude that L(Fn ) was a factor that all
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
9
nontrivial conjugacy classes are infinite; we call a group with this property
ICC (for infinite conjugacy classes). We have thus established:
Proposition 2.7. If Γ is an ICC group, then L(Γ) is a factor.
Another example of a factor is B(H) for any Hilbert space H.
Fact 2.8. Every von Neumann algebra is a direct integral of factors. It is
in this sense that the factors are the “prime” or “indecomposable” objects in
von Neumann algebra theory and one often proves facts about arbitrary von
Neumann algebras by first proving the result for factors.
We end this section with one of the most difficult open problems in von
Neumann algebra theory:
Question 2.9. If m, n ≥ 2 are distinct, is L(Fm ) ∼
= L(Fn )?
Even though Fm and Fn are clearly not isomorphic (nor are their group
rings for that matter), the group von Neumann algebra includes so many new
elements (as belonging to the weak closure is, well, a weak condition) that
it becomes much harder to distinguish things at the von Neumann algebra
level.
2.3. The Hyperfinite II1 factor R. In this subsection, we introduce arguably the most important von Neumann algebra. Towards this end, it will
be important to recall the Kronecker product of matrices. Suppose that A
is an m × n matrix and B is a p × q matrix. Then the Kronecker product of
A and B is the mp × nq block matrix


a11 B · · · a1n B

..
..  .
A ⊗ B :=  ...
. 
.
an1 B · · · ann B
We now consider the chain of inclusions
M2 (C) ,→ M4 (C) ,→ M8 (C) ,→ · · · ,
where S
the inclusion M2n (C) ,→ M2n+1 (C) is given by B 7→ B ⊗ I. We set
M := n M2n (C). The hyperfinite II1 factor R will be the von Neumann
algebra generated by M once we can view M as a set of operators on some
Hilbert space.
For each n, let trn : Mn (C) → C denote the normalized trace, namely
trn := n1 tr, where tr is the usual trace on matrices (so tr(I) = 1). Notice
that tr2n (B) = tr2n+1 (B ⊗ I) (this is the reason we work with normalized
traces), so that we get a unique function tr : M → C extending the individual
tr2n ’s. We can now define an inner product on M by hA, Bi := tr(B ∗ A).
(This is the extension of the usual Frobenius inner product on the space of
matrices.)
Let H be the completion of M (with respect to the norm inherited from
the aforementioned inner product). Then H is a Hilbert space and we can
view elements of M as operators on H. Indeed, given A ∈ M , the action of
10
ISAAC GOLDBRING
left multiplication by A, B 7→ A·B : M → M , extends to a bounded operator
on H. (Exercise) In this way, we get an algebra embedding M ,→ B(H) and
we let R be the von Neumann algebra generated by the (image) of M .
Exercise 2.10. R is a factor.
The adjective “II1 ” in the name of R will be defined later in these notes.
The word “hyperfinite” refers to the fact that R is the von Neumann algebra
generated by the increasing union of a countable chain of finite-dimensional
∗-subalgebras. It is a theorem of Murray and von Neumann that R is the
unique II1 factor with these properties. It follows that if Γ is an infinite ICC
group which is the increasing union of a countable chain of finite subgroups,
then L(Γ) is isomorphic to R. An example of a group with this property is
fin (N), which is the group of permutations of N with finite support.
S∞
We will come to see that R plays a special role in the theory of von
Neumann algebras (both from the operator algebra perspective as well as
the model-theoretic perspective).
3. Projections, Type Classification, and Traces
3.1. Projections and the spectral theorem. Recall that a bounded operator P : H → H is called a projection if P 2 = P ∗ = P . For example, if
P : H → H is the orthogonal projection onto a closed subspace of H, then
P is a projection. Conversely, if P is a projection, then P is the orthogonal
projection onto the closed subspace P (H) of H.
The goal of this subsection is to prove the following theorem:
Theorem 3.1. If A is a von Neumann algebra, then A is generated by the
projections in A.
It is for this reason that the study of von Neumann algebras relies on a
deep understanding of the projections in the algebra; we will discuss this in
the next subsection. For now, let us discuss the proof of Theorem 3.1. A
good reference for what is to follow is Conway’s book [4].
Recall that an operator N : H → H is normal if N N ∗ = N ∗ N . The
climax of a good undergraduate course in linear algebra is the following:
Theorem 3.2 (Spectral Theorem for Finite-Dimensional Operators). Suppose that dim(H) < ∞ and N : H → H is a normal operator. If λ1 , . . . , λk
are the distinct eigenvalues of N and Pi : H → H denotes the orthogonal
P
projection onto the eigenspace corresponding to λi , then N = ki=1 λi Pi .
Theorem 3.1 will follow from a suitable infinite-dimensional version of
the Spectral Theorem. In infinite-dimensional Hilbert spaces, rather than
discussing the eigenvalues of an operator, we will need to refer to its spectrum,
and instead of decomposing a normal operator as a sum of multiples of
projections, we will need to decompose it into an integral of such operators.
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
11
Definition 3.3. If X is a set, Ω is a σ-algebra of subsets of X, and H is a
Hilbert space, then a spectral measure for (X, Ω, H) is a function E : Ω →
B(H) such that:
(1)
(2)
(3)
(4)
For each C ∈ Ω, E(C) is a projection;
E(∅) = 0 and E(X) = I;
E(C1 ∩ C2 ) = E(C1 )E(C2 );
S
If
{C
:
n
∈
N}
are
pairwise
disjoint
elements
of
Ω,
then
E(
n
n Cn ) =
P
n E(Cn ) in the strong topology.
Suppose that E is a spectral measure for (X, Ω, H). Then for v, w ∈ H,
the function Ev,w (C) := hE(C)v, wi defines a countably additive measure
on Ω. Moreover, if f : RX → C is a bounded, Ω-measurable function, then
it is possible to define
R f dE, which
R is an element of B(H) satisfying the
important identity h( f dE)v, wi = f dEv,w for all v, w ∈ H.
For T ∈ B(H), recall that the spectrum of T is
σ(T ) := {λ ∈ C : (T − λI) is not invertible}.
Recall that σ(T ) is a nonempty, compact subset of C.
We can now state the Spectral Theorem.
Theorem 3.4 (Spectral Theorem). Suppose that N : H → H is a normal
operator. Then thereR is a unique spectral measure on the Borel subsets of
σ(N ) such that N = zdE.
An important by-product of the proof of the Spectral Theorem is the
following:
Theorem 3.5. Suppose that N : H → H is a normal operator and E
is the spectral measure for N . For T ∈ B(H), we have [T N = N T and
T N ∗ = N ∗ T ] if and only if T E(C) = E(C)T for every Borel subset C of
σ(N ).
Proof of Theorem 3.1. First, we notice that every element is a sum of two
∗
and =(T ) :=
normal operators. Indeed, given T ∈ B(H), set <(T ) := T +T
2
T −T ∗
2i . Then T = <(T ) + i=(T ) and <(T ) and =(T ) are normal (in fact,
self-adjoint). Moreover, if T ∈ A, then so are <(T ) and =(T ), whence A
is generated by its normal elements. Thus, it suffices to show that every
normal element of A is in the weak closure of the algebra generated by the
projections in A.
Suppose that N is a normal element of A. By Theorem 3.5, we see that the
spectral projections E(C) of N lie in the von Neumann algebra generated by
N and thus in A. It thus suffices to show that N is in the weak closure of the
algebra
generatedRby its spectral projections.
Fix v, w ∈ H. Then hN v, wi =
R
R
h( zdE)v, wi = zdE
.
Now
zdE
is
by the integral
v,w
v,w
P
P approximatedP
of a simple function
ci χCi , which is
ci Ev,w (Ci ) = h ci E(Ci )v, wi,
finishing the proof.
12
ISAAC GOLDBRING
3.2. Type classification of factors. For the rest of these notes, we keep
things simple and assume that all Hilbert spaces are assumed to have dimension ≤ ℵ0 .
Notation: We are going to switch from denoting elements of von Neumann
algebras by uppercase letters T and P and rather start using lowercase letters
x and p as we are now thinking of them as elements rather than operators.
If u ∈ B(H), we say that u is a partial isometry if u∗ u and uu∗ are
projections; in this case we call u∗ u and uu∗ the support projections and
range projections respectively and we call the spaces (u∗ u)(H) and (uu∗ )(H)
the initial space and final space respectively.
The next exercise helps explain the terminology.
Exercise 3.6. If u ∈ B(H) is a partial isometry with initial space H0 and
final space H1 , then u = u0 P , where P : H → H is the orthogonal projection
onto H0 and u0 : H0 → H1 is an (isometric) isomorphism of H0 onto H1 .
For the remander of this subsection, M denotes a von Neumann algebra.
Set P (M ) to be the set of projections in M . For p, q ∈ P (M ), we write p ≤ q
to mean p(H) ⊆ q(H).
Definition 3.7. For p, q ∈ P (M ) we say that p and q are Murray-von
Neumann equivalent, denoted p ∼ q, if there is a partial isometry u ∈ M
such that u∗ u = p and uu∗ = q.
Thus p and q being Murray-von Neumann equivalent means that the
spaces p(H) and q(H) are isomorphic and M knows that they are isomorphic.
We write p q to mean p ∼ p0 for some p0 ≤ q.
Exercise 3.8 (Murray-von Neumann Schröder-Bernstein). If p q and
q p, then p ∼ q.
Exercise 3.9. induces a partial order on P (M )/ ∼.
Example 3.10. If M = B(H), then p ∼ q if and only if dim(p) = dim(q) as
B(H) knows about all isomorphisms. Thus, defines a linear ordering on
P (B(H))/ ∼, which is order isomorphic to:
• {0, 1, . . . , n} if dim(H) = n, or
• ω + 1 if dim(H) = ℵ0 .
The fact that was a linear ordering on P (B(H))/ ∼ was not an accident:
Fact 3.11. M is a factor if and only if is a linear order on P (M )/ ∼.
Definition 3.12. Suppose that p ∈ P (M ). Then p is called:
• finite if p is not equivalent to a proper subprojection;
• infinite if it is not finite;
• purely infinite if it has no nonzero finite subprojection;
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
13
• semifinite if it is infinite but is the supremum of an increasing family
of finite subprojections;
• minimal if it is nonzero but has no proper nonzero subprojection.
For any of the adjectives above, we say that M is if I is (viewed as an
element of P (M )).
Example 3.13. B(H) is finite if dim(H) < ∞ and is otherwise semifinite.
(Remember that we are assuming that H has dimension at most ℵ0 .)
Which of the above adjectives applies to L(Γ)? We will soon see the
answer to that.
Definition 3.14 (Factor classification). Suppose that M is a factor. We say
that M is of type
• In if M is finite and P (M )/ ∼ is order isomorphic to {0, 1, . . . , n}
for n ∈ N;
• I∞ if M is infinite and P (M )/ ∼ is order isomorphic to ω + 1;
• II1 if M is finite and P (M )/ ∼ is order isomorphic to [0, 1];
• II∞ if M is semifinite and P (M )/ ∼ is order isomorphic to [0, ∞];
• III if M is purely infinite and P (M ) ∼ is order isomorphic to {0, ∞}.
Of course [0, 1] and [0, ∞ are order isomorphic; the point of writing them
this way is to indicate that I is finite in the former case and infinite in the
latter case. The same remark applies to {0, 1} and {0, ∞}.
Facts 3.15.
(1) Every factor is of one of the above types.
(2) There are factors of each type.
(3) Every type In factor is isomorphic to Mn (C).
(4) Every type I∞ factor is isomorphic to B(H), for dim(H) = ℵ0 .
(5) Classification of type II and type III factors is very difficult. For example, the isomorphism problem for either type II or type III factors
is not classifiable by countable structures (in the sense of descriptive
set theory).
We will be primarily concerned with type II1 factors. In fact, there will
be an excellent model-theoretic reason for doing so, as we will see later on
in these notes.
3.3. Traces. Throughout, M denotes a von Neumann algebra.
Definition 3.16. A (faithful, normal) trace on M is a function τ : M → C
satisfying:
• τ is a linear map;
• τ is positive, that is, τ (x∗ x) ≥ 0 for each x ∈ M ;
• (normality) τ is weakly continuous;
• (faithful) τ (x∗ x) = 0 if and only if x = 0;
• (trace property) τ (xy) = τ (yx).
14
ISAAC GOLDBRING
It often becomes convenient to normalize the trace and assume that τ (1) =
1.
Example 3.17. If dim(H) = n, it is not hard to see that every trace on
B(H) is a scalar multiple of the usual trace. If dim(H) = ℵ0 , then there is
no trace on B(H). Indeed, towards a contradiction, suppose that τ was a
trace on B(H). Let pn denote the orthogonal projection onto the subspace
spanned by the first n elements of some fixed orthogonal basis for H and
let qn denote the orthogonal projection onto the subspace spanned by the
nth element of the basis (so pn = q1 + · · · qn ). By faithfulness, we see that
τ (qi ) > 0 for each i. Since qi ∼ qj for all i, j, we will shortly see (Lemma 3.21)
that τ (qi ) = τ (qj ) for all i, j, whence τ (pn ) = nτ (q1 ). Since pn converges
strongly (and hence weakly) to I, normality would imply that τ (pn ) → τ (I);
but τ (pn ) → ∞, a contradiction.
Fact 3.18. A factor M is of type II1 if and only if it is infinite-dimensional
and posseses a trace, which is then necessarily unique if one assumes the
trace to be normalized.
Example
3.19. Recall that the hyperfinite II1 factor R was the completion
S
of n M2n (C) with respect to an appropriate inner
S product. Recall that
the inner product stemmed from a function tr : n M2n (C) → C which
was the extension of the normalized trace on each M2n (C). It is relatively
straightforward to show that tr extends to a trace on R, showing that R is
a II1 factor (whence the name is indeed appropriate).
Example 3.20. Let’s consider L(Γ), where Γ is an ICC group. We claim
that L(Γ) is a II1 factor. Since L(Γ) is certainly infinite-dimensional, it
remains to find a trace on L(Γ). Recall that we were thinking of elements of
L(Γ) as infinite pseudodiagonal matrices with respect to the standard basis
on `2 (Γ). If Γ were finite, then the normalized trace of such a pseudodiagonal
matrix would be the complex number that appears along the actual diagonal
of the matrix; thinking of the element of L(Γ) as a formal sum, it would be
thePcoefficient of uid . This suggests we define a function τ : L(Γ) → C by
τ ( γ∈Γ cγ uγ ) := cid ; alternatively, thinking of an element of L(Γ) as an
operator T on `2 (Γ), we define τ (T ) := hT id , id i. It is readily verified that
τ is a trace on L(Γ), whence we see that L(Γ) is a II1 factor. In particular,
this answers the question from before, namely that L(Γ) is a finite factor (in
the sense of Definition 3.12).
Traces are useful in factors because they detect Murray-von Neumann
equivalence:
Lemma 3.21. If M is a factor and τ : M → C is a positive, faithful linear
functional that satisfies the trace property, then for p, q ∈ P (M ), we have
p ∼ q if and only if τ (p) = τ (q).
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
15
Proof. First suppose that p ∼ q. Then there exists u ∈ M such that p = u∗ u
and q = uu∗ . But then
τ (p) = τ (u∗ u) = τ (uu∗ ) = τ (q).
Conversely, suppose that tr(p) = tr(q). Since M is a factor, we may assume
(by switching the roles of p and q if necessary) that p q. Then there is
p0 ∈ P (M ) such that p ∼ p0 and p0 ≤ q. Note that tr(p0 ) = tr(p) = tr(q) by
the first part of the proof. Meanwhile, we have q = p0 + q(1 − p0 ), whence
we have tr(q) = tr(p0 ) + tr(q(1 − p0 )) and consequently tr(q(1 − p0 )) = 0.
Since q(1 − p0 ) is a projection, we have q(1 − p0 ) = (q(1 − p0 ))∗ (q(1 − p0 )).
By faithfulness, we see that q(1 − p0 ) = 0, that is p0 = q, whence p ∼ q. We can think of traces as a dimension-type function in factors. However,
in II1 factors, {tr(p) : p ∈ P (M )} = [0, tr(1)], whence the dimension takes
on a continuum number of values.
The following fact will be utilized repeatedly later on in these notes:
Fact 3.22. R embeds into any II1 factor.
Proof. (Sketch) Let M be a II1 factor; we will show how to find a copy of
Mn (C) inside of M . (An elaboration on this idea allows one to embed all of
R into M .) First choose p1 , . . . , pn ∈ P (M ) of trace n1 that sum up to I; this
is possible because M is a II1 factor. Since p1 ∼ pj for each j ∈ {1, . . . , n},
∗ v = p and v v ∗ = p .
there are partial isometries v1j in M such that v1j
1j
1
1j 1j
j
∗ v . It is now an exercise to see that the von
For i ∈ {2, . . . , n}, set vij := v1i
1j
Neumann alebra algebra generated by {vij : 1 ≤ i, j ≤ n} is isomorphic to
Mn (C).
4. Tracial ultrapowers and the Connes Embedding Problem
4.1. The 2-norm. From now on, by a tracial von Neumann algebra we will
mean a von Neumann algebra A equipped with a trace τ . We will also
assume that the trace is normalized so that τ (1) = 1. We will often abuse
notation and just write A for a tracial von Neumann algebra and suppress
mention of the trace. (If A is a II1 factor, then this trace is unique and so
there is no loss of information in this notation.)
In the rest of these notes, we will be exclusively concerned with tracial
von Neumann algebras. Suppose that (A, τ ) is a tracial von Neumann algebra. We then get an inner product on A given by hx, yiτ := τ (y ∗ x) (this
is reminiscent of the inner product we considered when discussing R) which
then induces
pa norm on A, called the 2-norm on A and denoted k · k2 , that
is, kxk2 := τ (x∗ x).
Unlike the operator norm on A, the 2-norm is not submultiplicative, that
is, kxyk ≤ kxk · kyk but the corresponding fact need not hold for the 2-norm.
Nevertheless, the following inequality is often useful:
Lemma 4.1. For any x, y ∈ A, we have kxyk2 ≤ kxk · kyk2 .
16
ISAAC GOLDBRING
Proof. We will need two standard bits of functional analysis. The first is
that in any C ∗ algebra (and particular in a von Neumann algebra), we have
kyy ∗ k = kyk2 . Secondly, for any w ∈ A, we have w ≤ kwk · I (this follows
from the functional calculus); consequently, if x ∈ A, then wx∗ x ≤ kwkx∗ x
since x∗ x is a positive element. We are now ready:
kxyk22 = τ (y ∗ x∗ xy) = τ (yy ∗ x∗ x) ≤ τ (kyy ∗ kx∗ x) = kyk2 τ (x∗ x) = kxk22 ·kyk2 .
Note that the second equality follows from the trace property while the
inequality follows from positivity and linearity.
In particular, by setting y = I, we have that kxk2 ≤ kxk for all x ∈ A.
Fact 4.2. A is not complete with respect to the 2-norm. However, any
operator norm closed and bounded subset of A is complete with respect to
the 2-norm.
We call a tracial von Neumann algebra separable if A is separable with
respect to the metric induced by the 2-norm. For example, R and L(Γ) are
separable.
4.2. Tracial ultraproducts. Operator algebraists like taking ultraproducts/ultrapowers (almost) as much as model theorists do. Here is the definition of the tracial ultraproduct.
Definition 4.3. Suppose that (An : n ∈ N) is a sequence of tracial von
Neumann algebras and U is a nonprincipal ultrafilter on N. Set
Y
F((An )) := {(an ) ∈
An : sup kan k < ∞},
n
and
I((An )) := {(an ) ∈
Y
An : lim kan k2 → 0}.
U
F and I are to remind us of the words finite and infinitesimal. It is straightforward to check that F((An )) is an algebra and I((An )) is an operator
norm closed two-sided ideal. The quotient F((An ))/I((An )) is referred
Q to
as the tracial ultraproduct of (An ) with respect to U and is denoted U An .
If An = A for each n, then we refer to the tracial ultraproduct as the tracial
ultrapower of A and denote it by AU .
Remarks 4.4.
(1) Recall that if (rn ) is a bounded sequence of real numbers, then
limU rn = r means, for every > 0, we have {n ∈ N : |rn − r| < } ∈
U. It is a standard fact that the ultralimit of a bounded sequence of
real numbers always exist.
Q
(2) It is relatively straightforward to show that U An is once again a
tracial von Neumann algebra whose trace is given by tr = limU trn .
Less
Q trivially, if each An is a factor (resp. II1 factor), then so is
U An . This will also follow from the ability to axiomatize these
concepts in (continuous) first-order logic; see Subsection 5.1.
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
17
(3) The uses of k·k in F((An )) and k·k2 in I((An )) is not a typo. Indeed,
if one replaced k · k2 with k · k in the definition of I((An )), then one
would be performing the C ∗ ultraproduct; unless one is in a trivial
situation, the C ∗ ultraproduct of a sequence of von Neumann algebras is not a von Neumann algebra again. (See [13] for a wonderful
discussion of this issue.)
Remark 4.5. There are two bits of notational nuances that model theorists
should be aware of. First of all, operator algebraists like to use the notation
βN \ N to denote the set of all nonprincipal ultrafilters on N. Secondly,
operator algebraists like to use the notation ω for an element of βN \ N; this
clearly causes confusion for logicians and we will refrain from this practice.
We will see later that the tracial ultraproduct is precisely the modeltheoretic ultraproduct in an appropriate continuous logic.
4.3. The Connes Embedding Problem. In 1976, Connes proved the following result:
Theorem 4.6. Fix U ∈ βN \ N. Then L(Fn ) embeds into RU .
He then remarked that the previous fact “ought to be true for any separable II1 factor.” This remark has yet to be proven and the resulting problem
is known as the Connes Embedding Problem (note the use of the word “problem” rather than “conjecture”). It is arguably the most important unsolved
problem in operator algebras. It has an unbelievable number of surprising
equivalent formulations (see [3]). It also has an interesting interaction with
the model theory of tracial von Neumann algebras, as we will see a bit later.
Let us introduce some terminology that will be useful. Once the modeltheoretic appartus is set up in the next section, we will see that, for U, V ∈
βN \ N and A a separable tracial von Neumann algebra, A embeds into
RU if and only if A embeds into RV . For that reason, we will say that
A is Rω -embeddable if A embeds into some (equivalently any) nonprincipal
ultrapower of R. (Yes, I know, I am acquiescing to the operator algebraist
notation.) Thus, the Connes Embedding Problem (which we will abbreviate
CEP) asks whether every separable II1 factor is Rω -embeddable.
Fact 4.7. Any tracial von Neumann algebra embeds into a II1 factor. Indeed, if A is a tracial von Neumann algebra, then A ∗ L(Z) (free product) is
a II1 factor and A ⊆ A ∗ L(Z).
Thus, we may replace the word “II1 factor” by “tracial von Neumann algebra” in the statement of the CEP.
5. Some Model Theory of tracial von Neumann algebras
In this section, we will highlight some of the most striking aspects of the
model theory of tracial von Neumann algebras. We should remark that we
are still at the early stages of our understanding of the situation. We will
18
ISAAC GOLDBRING
be brief with our discussion, referring the reader to the appropriate parts of
the literature where these things are discussed in greater detail.
5.1. Axiomatizability. The first task is to make the tracial von Neumann
algebras the models of a theory in an appropriate logic. The approach developed in [7] is to work in a version of continuous logic (a la [1]) that contains
“domains of quantification.” The metric on a structure is required to be complete with respect to each domain of quantification and the language has to
specify a modulus of uniform continuity for the symbols on each domain.
For tracial von Neumann algebras, the metric is the one induced by the
2-norm. Since the 2-norm is complete with respect to operator norm closed
and bounded sets, the domains of quantification Dn correspond to the ball of
radius n around 0 with respect to the operator norm. We include a symbol
for the 2-norm, but not the operator norm. Indeed, the operator norm is not
uniformly continuous with respect to the metric induced by the 2-norm and
could thus not be asked to be a distinguished predicate.
However, writing down a set of axioms that captures all of the structures
that are the result of viewing tracial von Neumann algebras in this way is
tricky business. Indeed, while asking that the structure is a tracial unital
∗-algebra is not so difficult, requiring that the domains Dn correspond to
the operator norm balls is difficult considering that one is not allowed to
refer to the operator norm in the axioms! Nevertheless, it is shown in [7]
that one can axiomatize the class of structures that are the result of viewing
tracial von Neumann algebras in the way specified above. Since the class
of tracial von Neumann algebras are closed under subalgebras, the resulting
axiomatization should be universal. (This was not the case in the original
version of [7].) We will henceforth refer to the theory of tracial von Neumann
algebras in this logic as TvNa .
We should mention that verifying the correctness of the axioms for tracial
von Neumann algebras is nontrivial and uses some facts from von Neumann
algebra theory, including functional calculus, the Kaplansky density theorem,
the Russo-Dye theorem, and the GNS construction.
One feature of the logic used to study tracial von Neumann algebras is
that the corresponding ultraproduct construction, when applied to tracial
von Neumann algebras, is precisely the tracial ultraproduct construction.
This yields a proof of the fact that the tracial ultraproduct of a family of
tracial von Neumann algebras is once again a tracial von Neumann algebra.
Since the weak closure of a union of a chain of factors is once again a
factor, the class of factors should be ∀∃-axiomatizable. An explicit set of
∀∃-axioms is given in [7]; proving the correctness of these axioms requires
the Dixmier Averaging Theorem. Once again, this yields as a corollary that
the tracial ultraproduct of a family of factors is once again a factor.
Similarly, the class of II1 factors is also ∀∃-axiomatizable. The axiom one
needs to add to the axioms for factors in order to get the axioms for II1
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
19
factors is quite easy to write down. In fact, here it is:
1
inf (kxx∗ − (xx∗ )2 k2 + | tr(xx∗ ) − |) = 0.
x∈D1
π
If M is a II1 factor, then by the remark following Lemma 3.21, we know that
there is a projection in M of trace π1 , whence this projection witnesses that
M satisfies this axiom. On the other hand, suppose that M is a tracial factor
that satisfies the above axiom but is not a II1 factor. From our discussion
in Section 3.3, we see that M is thus a type In factor for some n, whence
isomorphic to Mn (C) for some n. Since D1 (Mn (C)) is compact, the inf
axiom is actually realized rather than approximately realized, say by a ∈
D1 (Mn (C)). Let p = aa∗ ; then p is a projection of trace π1 . However, the
traces of projections in Mn (C) are of the form nk for k ∈ {0, 1, . . . , n}, a
contradiction. (We thus see that the actual value of π1 was not relevant, only
that it was irrational.) Once again, we see that the tracial ultraproduct of a
family of II1 factors is once again a II1 factor.
The fact that the class of II1 factors is ∀∃-axiomatizable has a very interesting model theoretic consequence. Recall that a model M of a theory
T is an existentially closed model of T if for any extension N ⊇ M that is
also a model of T and any formula ϕ(x) with parameters from M , we have
(inf x ϕ(x))M = (inf x ϕ(x))N . Since any model of TvNa extends to a II1 factor
(Lemma 4.7) and the theory of II1 factors is ∀∃-axiomatizable, we see that
Proposition 5.1. Any existentially closed tracial von Neumann algebra is
a II1 factor.
This proposition was the model theoretic reason for appreciating II1 factors alluded to at the end of Subsection 3.2.
We should mention that we only know of one concrete example of a separable existentially closed II1 factor, namely R, although we know that there
are continuum many separable existentially closed II1 factors; see [5].
While we are on the topic of axiomatizability, let us discuss one interesting
side note. In order to show that R and L(F2 ) are not isomorphic, Murray and
von Neumann isolated a property of II1 factors that R has and that L(F2 )
does not. The property, called property (Γ), says that, for every finite tuple
x̄ and every > 0, there is a trace 0 unitary u such that [xi , u] := xi u − uxi
has 2-norm less than . It turns out that property (Γ) is axiomatizable in
continuous first-order logic. Indeed, for each n ≥ 1, let σn be the sentence
n
X
∗
sup inf (ky y − Ik2 + |τ (y)| +
k[xi , y]k2 ),
x̄
y
i=1
σnM
where x̄ is an n-tuple. Notice that
= 0 does not at first glance guarantee that there exists a trace 0 unitary that almost commutes with each xi ,
but that there is an “almost” unitary of small trace that almost commutes
with each xi (as inf’s are not necessarily realized in an arbitrary structure);
however, a standard functional calculus trick allows one to move that near
20
ISAAC GOLDBRING
witness to an actual trace 0 unitary with the desired property. It follows
that a II1 factor satisfies property (Γ) if and only if it makes each σn equal
to 0. As a consequence, we see that R 6≡ L(F2 ).
The following weaker version of Question 2.9 is still open:
Question 5.2. If m, n ≥ 2 are distinct, is L(Fm ) ≡ L(Fn )?
5.2. The model theoretic version of CEP. Suppose that A is an Rω embeddable tracial von Neumann algebra. Then, by standard model theory,
we have that A |= Th∀ (R). Conversely, suppose that A |= Th∀ (R) is separable. Let M |= Th(R) be separable such that A embeds into M ; we may
choose M separable by Downward Löwenheim-Skolem. Then since RU is an
ℵ1 -saturated model of Th(R), we have that M embeds into RU , whence A
is Rω -embeddable. In other words:
Lemma 5.3. If A |= TvNa is separable, then A is Rω -embeddable if and only
if A |= Th∀ (R).
Since TvNa is universally axiomatizable, we see that:
Corollary 5.4. CEP is equivalent to the statement TvNa = Th∀ (R).
We should mention that, using “abstract model-theoretic nonsense,” Hart,
Farah, and Sherman proved in [8] that there exists a separable II1 factor
S such that TvNa = Th∀ (S); they call such an S locally universal. In fact,
once there exists one locally universal II1 factor, then there exists many
locally universal II1 factors, e.g. if S is locally universal, so is S ∗ A for any
A |= TvNa . CEP asks whether or not R is locally universal.
5.3. Model companions and connection to CEP. Based on work of
Nate Brown [2], the following appears in [9]:
Theorem 5.5. Th(R) does not have quantifier elimination.
We also noticed that the proof of the preceding theorem applied to a
wider class of von Neumann algebras. First, we need a definition. Call a II1
factor A McDuff if A ⊗ R ∼
= R. (We have not defined the tensor product
of von Neumann algebras, but it works as one might guess; see [10] for more
details.) For example, R is McDuff.NIndeed, a fancier way of describing
our construction of R was that R = ∞
n=1 M2 (C); it seems quite plausible
(and is in fact the truth) that the tensor product of two infinite tensor
powers of M2 (C) would be isomorphic to one such infinite tensor power.
More generally, if A is a II1 factor, then A embeds into a McDuff II1 factor,
namely A ⊗ R.
Theorem 5.6. If S is locally universal and McDuff, then Th(S) does not
have quantifier elimination.
In [8], it is shown that, for separable II1 factors, being McDuff is an ∀∃axiomatizable property. Since every II1 factor embeds into a McDuff II1
factor (by tensoring with R), it follows that every existentially closed II1
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
21
factor is McDuff. (On a side note, if S is locally universal, so are S ⊗ R and
S ∗ R; since the first is McDuff and the second is not, we see that not all
locally universal II1 factors are elementarily equivalent.)
The previous observation, combined with the fact that TvNa has the amalgamation property, yields the following negative result:
Theorem 5.7 (G., Hart, Sinclair [9]). TvNa does not have a model companion.
Proof. Suppose that Th(S) is the model companion of TvNa , where S is separable. Since Th(S) is model complete, it follows that S is existentially closed,
whence a McDuff II1 factor. Moreover, since TvNa = Th∀ (S), we see that
S is locally universal. On the other hand, since TvNa has the amalgamation
property, it follows that the model companion has quantifier elimination, a
contradiction.
Instead of asking for a model companion of TvNa , one might ask for a model
complete theory of II1 factors. This question has an interesting connection
to CEP. First, two preparatory results.
Fact 5.8. Every embedding R → RU is unitarily conjugate to the diagonal
embedding. In particular, every embedding R → RU is elementary.
Exercise 5.9. Use Fact 5.8 to prove that R is the prime model of its theory.
Lemma 5.10 ([9]). If M is an Rω -embeddable II1 factor such that Th(M ) is
∀∃-axiomatizable (in particular, if Th(M ) is model-complete), then M ≡ R.
Proof. (Sketch) This is a “sandwiching chain” argument. To begin the argument, one embeds R into M (Fact 3.22) and M into RU and uses the
fact that the composition of these embeddings is elementary by Fact 5.8.
One continues the chain by taking the ultrapowers of the aforementioned
maps.
Thus, the only possible ∀∃-axiomatizable complete theory of II1 factors is
Th(R). It is still currently unknown whether or not Th(R) is ∀∃-axiomatizable.
Theorem 5.11. If CEP has a positive solution, then there is no model complete theory of II1 factors.
Proof. Suppose that Th(M ) is model complete, where M is a II1 factor. By
CEP, M is Rω -embeddable, whence, by Lemma 5.10, Th(M ) = Th(R). By
CEP again, we see that Th(R) is the model companion of TvNa , contradicting
Theorem 5.7.
A more careful analysis led to the following sharpening of the Theorem
5.11:
Theorem 5.12 (Farah, G., Hart [5]). If Th∀ (R) has the amalgamation
property, then there is no model complete theory of II1 factors.
22
ISAAC GOLDBRING
The previous result is indeed sharper than Theorem 5.11 as the fact that
TvNa has the amalgamation property shows that CEP implies that Th∀ (R)
has the amalgamation property. In fact, in [5], we exhibit a concrete Rω embeddable tracial von Neumann algebra such that the ability to amalgamate over it while staying Rω -embeddable yields the fact that there is no
model complete theory of II1 factors.
5.4. Stability. A major impetus for understanding the model theory of tracial von Neumann algebras stemmed from many questions of the form: “How
canonical are tracial ultraproducts and ultrapowers” For example, if M is a
separable tracial von Neumann algebra and U, V ∈ βN \ N, is M U ∼
= MV?
(There were many variants of this questions asked by many different people, but let us focus on this particular question.) Now, if the continuum
hypothesis holds, then M U and M V are saturated elementarily equivalent
structures of the same uncountable density character. Thus, a familiar back
and forth argument shows that they are isomorphic. Thus, the question is
only interesting if we assume the negation of the continuum hypothesis.
The following theorem (in its classical form) is model-theoretic folklore,
but a proof can be found in [7]:
Theorem 5.13. Suppose that the continuum hypothesis fails and M is a
separable structure in a countable language. Then the following are equivalent:
(1) For every U, V ∈ βN \ N, we have M U ∼
= MV.
(2) Th(M ) is stable.
A word about the proof: the direction (1) ⇒ (2) proceeds by showing
that a witness to the order property allows one to encode the partial order on increasing sequences of natural numbers (modulo almost everywhere
agreement) by almost everywhere domination into the structure and then
use the fact that there are nonprincipal ultrafilters whose associated partial
orders have distinct coinitalities (a result due independently to Dow and
Shelah). The other direction proceeds by showing that stability still allows
one to conclude that M U is saturated: one realizes a type over a large set by
finding Morley sequences for its restriction to a countable set of size continuum in M U and then using some forking calculus and statonarity of types
to find an actual realization of the original type.
Suppose that M is a II1 factor. By Fact 3.22, one can find arbitrarily
large matrix algebras in M . A straightforward computation with matrices
allows one to find an order in M , whence II1 factors are unstable and hence
have nonisomorphic ultrapowers (assuming the negation of the continuum
hypothesis). Similar reasoning allows one to show that there are nonisomorphic ultraproducts of matrix algebras. (Details for these two facts can be
found in [6].)
Which tracial von Neumann algebras are stable? That question is also
answered in [6]:
A GENTLE INTRODUCTION TO VON NEUMANN ALGEBRAS
23
Theorem 5.14 (Farah, Hart, Sherman). If M |= TvNa , then M is stable if
and only if it is of type I.
A word is in order about the statement of Theorem 5.14. We only defined
the type classification for factors, but there is also a type classification for
arbitrary von Neumann algebras. We will not explain this in detail, but only
say as much as is needed to understand the previous result. A tracial von
Neumann algebra M can be written M = MII1 ⊕ MI2 ⊕ MI3 ⊕ · · · , where
the subscript tells us the type of the algebra. A type In A algebra is of the
form Mn (C) ⊗ Z(A) and M as above is said to be of type I if its type II1
component is 0.
A word about the proof of Theorem 5.14. If M is not of type I1 , then
by our earlier discussion, M is unstable. It remains to show that type I
algebras are stable; equivalently, we show that all ultrapowers of type I algebras are isomorphic. It can be shown that ultrapowers and direct sums
“commute”, so it suffices to show that ultrapowers of type In algebras are isomorphic. It can be shown that (Mn (C) ⊗ Z(A))U ∼
= Mn (C)U ⊗ Z(A)U . Since
U
∼
(Mn (C)) = Mn (C) (compact things are isomorphic to their ultrapowers),
it suffices to show that the ultrapowers of abelian von Neumann algebras are
all isomorphic. There are many ways to explain this, but perhaps the easiest
is to recall that there is an equivalence of categories between the category of
abelian von Neumann algebras and the category of atomless probability algebras, respecting the ultraproduct construction; the latter theory is known
to be stable (see [1]), finishing the proof of the thoerem.
References
[1] I. Ben Yaacov, A. Berenstein, C. W. Henson, A. Usvyatsov, Model theory for metric
structures
[2] N. Brown, Topological dynamical systems associated to II1 factors, Adv. Math. 227
(2011), 1665-1699.
[3] V. Capraro, A survey on Connes’ embedding conjecture, arXiv 1003.2076.
[4] J. Conway, A course in functional analysis, GTM 96.
[5] I. Farah, I. Goldbring, B. Hart, Amalgamating, Rω -embeddable II1 factors, note available at http:///www.math.uic.edu/ isaac
[6] I. Farah, B. Hart, D. Sherman, Model theory of operator algebras I: Stability, to appear
in the Bulletin of the London Mathematical Society.
[7] I. Farah, B. Hart, D. Sherman, Model theory of operator algebras II: Model theory,
preprint.
[8] I. Farah, B. Hart, D. Sherman, Model theory of operator algebras III: Elementary
equivalence and II1 factors, preprint.
[9] I. Goldbring, B. Hart, T. Sinclair, The theory of tracial von Neumann algebras does
not have a model companion, to appear in the Journal of Symbolic Logic.
[10] V. Jones, Von Neumann algebras, lecture notes from his webpage.
[11] M. Lupini and A. Tornquist,
Set theory and von Neumann algebras,
lecture
notes
from
Appalachian
set
theory.
Available
at
http://www.math.cmu.edu/ eschimme/Appalachian/AsgerMartino.pdf
[12] J. Lurie,
Course notes on von Neumann algebras,
available at
http://www.math.harvard.edu/ lurie/
24
ISAAC GOLDBRING
[13] V. Pestov, Hyperlinear and sofic groups: a brief guide, Bull. Symb. Logic, 14, Number
4 (2008).
Department of Mathematics, Statistics, and Computer Science, University
of Illinois at Chicago, Science and Engineering Offices M/C 249, 851 S.
Morgan St., Chicago, IL, 60607-7045
E-mail address: [email protected]
URL: http://www.math.uic.edu/~isaac